Rasa

Overview

A clinical AI system that converts unstructured MRI doctor notes into a standardized structured format using Meta LLaMA, enabling faster comparison of current and previous patient findings.

The Problem

Doctors often write notes based on MRI reports after analyzing radiology images to keep accurate patient records. However, they need to compare current findings with previous notes, which makes the process time-consuming and difficult.

One major issue is the lack of a well-organized system to store and manage these notes. Because of this, retrieving and comparing past MRI records becomes inefficient and adds extra workload for doctors.

Challenges

  • No clinically validated benchmark datasets
  • Free-form variability in MRI doctor notes
  • Reliable extraction of structured medical variables
  • Building robust prompt-to-fine-tuning workflows

The Solution

Built SRI (Structured Radiology Intelligence) to transform unstructured MRI notes into 55 structured clinical variables, making historical comparison and downstream analysis significantly easier.

Dataset Innovation

Created a novel benchmark dataset based on 55 clinically relevant variables defined by doctors, generated through systematic prompt tuning and manually validated for real-world reliability.

Model Development

The system was developed in three stages:

  • Baseline - simple prompting with Meta LLaMA
  • Intermediate - prompt strategy optimization
  • Final - fine-tuning using the validated benchmark dataset

Each stage focused on improving structured extraction accuracy.

Final Outcome

  • Built a clinically relevant benchmark dataset
  • Fine-tuned Meta LLaMA for radiology extraction
  • Developed an end-to-end MRI note structuring pipeline
  • Improved report analysis efficiency for doctors

Tech Stack

  • Frontend: Next.js
  • Backend: FastAPI, PyTorch, Hugging Face
  • Languages: JavaScript, Python
  • Database: MongoDB